The paper introduces a new training strategy called Staged Depth Training (SDT) that implements a representation curriculum for physics-informed neural networks (PINNs). SDT trains a shallow prefix with a temporary physics-informed head, freezes it, and then adds depth, improving performance across many benchmark problems without changing the final architecture. Experiments show significant error reductions and deeper fitted depth-scaling exponents, indicating that explicit representation learning can enhance PINN performance.
arXiv:2606. 16112v1 Announce Type: cross Abstract: Residual architectures are ubiquitous in deep learning, but they suffer from a subtle structural limitation: the norm of the residual stream can grow rapidly with depth.
arXiv:2607.01630v2 Announce Type: replace Abstract: Dynamic expansion methods for class-incremental learning (CIL) protect task-specific knowledge by growing dedicated tokens or subnetworks, yet our...
arXiv:2608.21756v1 Announce Type: cross Abstract: Deep-learning efforts have increasingly shifted toward foundation model approaches. In experimental physics, this allows models and learned represent...
arXiv:2609.40127v1 Announce Type: cross Abstract: Modern transformers pair impressive capabilities with substantial memory and compute demands. Low-rank weight factorization reduces both while keepin...
Dynamic expansion methods for class-incremental learning (CIL) protect task-specific knowledge by growing dedicated tokens or subnetworks, yet our analyses suggest that classification supervision alone does not sufficiently preserve task-agnostic shared backbone representations over long incremental sequences. We identify two intertwined challenges: cross-task confusion from sequential training on predominantly current-task data, which biases decision boundaries toward recent tasks; and under-optimized shared representations in the backbone that cap long-term discriminability as tasks accumulate.
The paper investigates why latent neural surrogate solvers, which compress physical system dynamics into a lower‑dimensional space, often fail during long‑horizon autoregressive rollouts. It demonstrates that training the latent representation only for reconstruction leads to instability, and proposes a set of training interventions—Koopman operator learning, Hamming noise injection, and multi‑step rollout fine‑tuning—that align the latent space with long‑horizon forecasting. These interventions reduce long‑rollout error by about 40 % and achieve accuracy comparable to full‑resolution models while using far fewer floating‑point operations and GPU memory, enabling stable extrapolation in mesoscale crystal‑plasticity simulations of high‑cycle fatigue.
arXiv:2603. 19312v3 Announce Type: replace Abstract: Joint Embedding Predictive Architectures (JEPAs) offer a compelling framework for learning world models in compact latent spaces, yet existing methods remain fragile, relying on complex multi-term losses, exponential moving averages, pre-trained encoders, or auxiliary supervision to avoid representation collapse.
arXiv:2607. 20652v1 Announce Type: cross Abstract: Language models are thought to exhibit the phenomenon of superposition, representing many more features than dimensions in their residual streams.
arXiv:2607. 14018v1 Announce Type: cross Abstract: We investigate how each component of the Transformer feedforward block architecture design determines how much rank survives across depth at initialization.
arXiv:2606. 29223v1 Announce Type: new Abstract: Autoregressive LLM decoding evaluates every generated token through the full layer stack, even though many tokens become predictable at intermediate depths.
DART is a new RGB‑D pretraining method for surgical vision foundation models that incorporates pseudo‑labeled depth maps as a pixel‑space reconstruction target during training. By adding a depth reconstruction head to DINOv2’s masked iBOT framework, DART improves representation quality without affecting downstream RGB‑only fine‑tuning or inference. Across eight surgical benchmarks—including segmentation, depth estimation, and image‑level recognition—DART outperforms both natural‑image and in‑domain baselines, demonstrating that geometric pseudo‑labels can strengthen foundation model pretraining without extra labels or inference cost.
Modern transformers pair impressive capabilities with substantial memory and compute demands. Low-rank weight factorization reduces both while keeping the matrices dense, and thus efficient on standar...